Relative information extraction by in-context and skill learning

Determine whether In-Context Learning and Skill Learning extract meaningful information from agent experience in equivalent or substantively different ways when transferring skills from basic tasks to unseen advanced embodied tasks.

Background

The benchmark compares two experience-use strategies: In-Context Learning, which retains complete interaction trajectories, and Skill Learning, which supplies a concise skill summary generated after the learning phase. Neither strategy consistently outperforms the other across the evaluated harnesses and models.

The authors explicitly conjecture that both methods can extract meaningful information from experience, but the claim is not established by the reported results. Clarifying what information each representation extracts and how that information supports transfer remains an unresolved question.